Refs librefang/librefang#4842 — long-term replacement for the substring
match that the OpenAI driver currently uses to decide how to handle
`reasoning_content` on historical assistant turns.
Three provider-specific behaviours that the driver must distinguish at
wire time, now expressed as catalog metadata:
* `strip` — DeepSeek R1 / deepseek-reasoner. The API rejects requests
that carry reasoning_content on previous assistant messages.
* `echo` — DeepSeek V4 Flash. Thinking mode is on by default and the
API rejects multi-turn requests when assistant turns containing
tool_calls don't echo back the original reasoning text. This is the
bug surfaced in librefang/librefang#4842.
* `empty_string` — Moonshot / Kimi K2 family. The field must be present
(empty string) on tool_calls turns, with thinking disabled wire-side
for multi-turn compatibility.
* `none` (default) — most providers; field is omitted entirely.
V4 Pro is intentionally NOT marked `echo` — librefang#4842 reports it
working out-of-the-box; flip when there's an empirical reproducer.
Marks affected models:
providers/deepseek.toml
deepseek-v4-flash → echo
deepseek-reasoner → strip
providers/moonshot.toml
kimi-k2.6, kimi-k2.5, kimi-k2 → empty_string
providers/kimi-coding.toml
kimi-for-coding → empty_string
providers/byteplus-coding.toml
kimi-k2.5 → empty_string
providers/novita.toml
moonshotai/kimi-k2-thinking → empty_string
Tooling:
* schema.toml registers the field with the four enum options and a
`none` default so existing TOML files keep parsing unchanged.
* scripts/validate.py rejects unknown enum values; verified with a
hand-crafted negative case (`reasoning_echo_policy = "bogus"` →
validation fails with the expected message).
* `python3 scripts/validate.py` passes (267 models).
The librefang side that consumes this field will land in a follow-up
PR — until then, registry consumers ignore the field via
`#[serde(default)]` and the existing substring fallback continues to
work, so this commit is safe to ship independently.
* feat: add 4 context engine plugins
- topic-memory: keyword clustering for topic-aware memory recall
- episodic-memory: conversation segmentation and cross-session recall
- user-profile: persistent user profiling from conversation patterns
- context-decay: time-based memory decay with reinforcement dynamics
All plugins use the ingest/after_turn hook protocol with stdin/stdout JSON.
* chore: add plugin scaffolding, update docs and templates
- Add plugin.toml template with {{NAME}} placeholder
- Add new-plugin Makefile target with hooks/ scaffolding
- Update plugins/README.md with all 10 plugins
- Update README.md stats (10 plugins, 220+ models)
- Add Plugin checkbox and checklist to PR template
- Add Plugin to issue template content type dropdown
- Fix CONTRIBUTING.md: last_verified is recommended, not required
* fix: correct model pricing and remove deprecated entries
- openrouter/gemma-2-9b-it: fix pricing from 0.0 to 0.03/0.09 per M tokens
(free variant correctly stays at 0.0)
- github-copilot: remove deprecated copilot/gpt-4 model entry
(GPT-4 retired in favor of GPT-4o for Copilot)
* docs: annotate kimi-coding as membership-gated
Kimi Code CLI uses quota-based membership model (not per-token billing).
Free tier has limited weekly requests; underlying model is K2.5.
Pricing kept at 0.0 consistent with other subscription providers
(chatgpt, github-copilot) but with explanatory comments.
* style: fix trailing newline in github-copilot.toml
* fix: correct Moonshot/Kimi model pricing from official sources
All 5 models had incorrect pricing:
- moonshot-v1-8k: 0.10/0.10 → 0.20/2.00
- moonshot-v1-32k: 0.30/0.30 → 1.00/3.00
- moonshot-v1-128k: 0.80/0.80 → 2.00/5.00
- kimi-k2: 2.00/8.00 → 0.60/2.50
- kimi-k2.5: 2.00/8.00 → 0.45/2.20
Sources: platform.moonshot.ai/docs/pricing/chat, costgoat.com, getmaxim.ai
* feat: add MiniMax M2.7 and M2.7-highspeed models
Released 2026-03-18, MiniMax's latest flagship text model.
10B activated params, 200K context, 128K output, tool use, streaming.
Pricing: $0.30/$1.20 per M tokens (input/output).
Added to both international (minimax.io) and China (minimaxi.com) providers.
Community-maintained TOML catalog for LibreFang. New models can be added
via PR without requiring a LibreFang binary release.
Includes validation script, bilingual docs, and GitHub templates.
Evan